Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable

[printf42]

Just heard a report on Tesla autonomous event.

They are skeptical about the 2020 timeline. “Many autonomous projects have been pushed back recently, for example such and such... Will Tesla make it in 2020? Their event yesterday was late by half an hour!”

I actually don’t have any counter argument, LOL

 

[neroden]

If they really did all this in 3 months, isn’t it the biggest competitive advantage?
Imagine where they will be in 1 year with that rate of improvement.

It actually kind of is. Tesla is barking up all the wrong trees, one at a time, but at least in this area, they're doing it very fast. I'm watching for when they actually understand the problems they're dealing with, which is going to take a couple more years IMO. They're just BARELY starting to recognize the nature of the edge case problem, and they have vastly underestimated how hard it is.

However, the so-called "competition" seem to have underestimated the difficulty by even more. Simulations are not effective at finding the edge cases that cossetted California engineers had never imagined, as Karpathy repeatedly pointed out.

Tesla is the only company in a position to start working on true self-driving, because nobody else has the fleet to get the data. There's a hell of a lot to do after you have the fleet -- decades of work, in my opinion -- but nobody else is really in a position to get started, as far as I can tell.

You'd think a second-mover could just skip all of that and bark up the right tree immediately, but history shows they don't. They end up iterating through mistakes too. So this puts Tesla in a strong lead.

However, all cars will still have drivers, who have to be attentive, in 2020, and 2021, and 2022, and most likely 2023. I wish Musk would stop overpromising!

If there are board members who knew how awful this event would look from an investment point of view, that would explain why it was scheduled BEFORE q1 financial results.

 

[Remus]

If they only have one photo of a weird edge case

there is several example of this in the presentation. one of them is they found an edge case (a single picture) the car has trouble identifying lanes in a tunnel, then they just ask all the cars to look for tunnels where they have trouble with lane lines, when such criteria is hit, upload the clip to mothership.

 

[EVNow]

nah, I won't be so worried, software side it's a team of several dozen software engineers, say 250k per year compensation on average, partly in stock, with data collection cost paid to AT&T and computation cost as bread crumbs, 20 million dollars per year tops.

Yes - I was calculating using 300k a year for 200 people, so $60M. Ofcourse they are getting EAP out of it. If 60k cars buy EAP per year, that cost is covered. They are basically developing FSD for free.

 

[jerry33]

Um, since when has Tesla only shown two executives (with no titles) on its Management page?

No CFO, no other C-level names listed. That's new.

https://ir.tesla.com/corporate-governance/management

View attachment 399835

Michelin used to do that. Maybe they still do. Everyone's title was "manager".

 

[Bladerskb]

I think you missed a few important things about Tesla's design goals:

1. Performance at a batch size of 1. The GPU solution nVidia sells requires data be collected, and processed in batches. The TOPs measure is a measure of throughput, but Tesla wanted throughput and super low latency so they went with a design that achieve's it's performance at a batch size of 1. Low latency is a critical differentiator when it comes to the automotive environment and safety.

2. Power: Pegasus operates in a 5x power envelope, which besides using up more range, also requires a lot more cooling and will have higher reliability risks because of the heat and cooling system required to remove it.

3 Cost, I don't know how much Pegasus costs, but we know that HW3 costs 80% of what the HW 2.5 chipset does. In the Autonomy day webcast they explain how they basically paid for the development by achieving 80% cost, so that over time the money spent on development of the chip will be paid back in money saved per unit. A very smart move.

I don't see Pegasus as an attractive solution to OEMs selling consumer vehicles, it's too big, hot and expensive. Seems the main audience is fleet operators. But it should be clear that the future is dedicated ASIC's like Tesla's (And Mobileye's) design.

This is incorrect. Again I'm only pointing out things that are clearly incorrect. I need every Tesla fan who have the absolute truth to their disposal. Nvidia performance at batch size of 1 has exponentially increased compared to Drive Px 2. Again Tesla is comparing a 3 year old, 3 gen old chip to a chip from 2019 rather than comparing a chip from 2019 to a chip from 2019. They also compared their chip to the TPU1 when talking about "batch size of 1". But notice that its TPU1 which came out 4-5 years ago.

Again you have to look at every Tesla comparison in close context. The devil is in the detail.

 

[Wnuk]

Don't think this video from Tesla has been posted here yet; quite impressive.

Tesla on Twitter

 

[NakedLong]

Agree with all of this except for one thing. Based on similar problems, I could estimate a time frame. Decades.

FSD is essentially a solved problem and now its just a matter of extending the training over the next 2-3 years. Sorry Neroden, but I'm going to trust what Karpathy says over you.

 

[Anstandswauwau]

Don't think this video from Tesla has been posted here yet; quite impressive.

Tesla on Twitter

It was shown during yesterday's presentation.

 

[NicoV]

I don't like the use of this term either. Most people will interpret it as "we're done and ready for customers" when it more likely means "we're out of development mode and moving into testing mode now and it will take a while for it to be ready for customers"

Fyi: in software development, ‘feature complete’ means all features have been developed, but quality may not be good enough. ‘Done and ready for customers’ is ‘general availability’. In between those stages you have alpha and beta software, with various completeness and quality levels.

 

[EVNow]

Obeying verbal instructions from police is a (relatively) trivial subset of natural language processing. That's like saying you need to have a machine shop in order to turn a screw. Won't quibble with the rest

Most of the time the command is actually not just verbal - it also involves kind of sign language (afterall everyone may not understand the language). Easy enough to figure out. Even a cheap kinect can do it.

 

[dc_h]

Sorry I disagree with you here. You are focused on individual corner cases, not a set of them. Guess what? When you a million car fleet, all those corner conditions will end up with hundreds of data points.

They will assume that of those hundred data points for a given edge condition, the vast majority will have been handled properly. Since the model should not be overfit anyways, its should generalize performance over these hundred of cases. If >90% of the automatic labels are good, this should converge properly.

TLDR you are obsessed with methods that matter when you have small amounts of data, but Tesla will have a metric s**t ton of it. Different story.

This is why they’ll start in San Francisco. Besides social acceptance, the neural net density will eliminate corner cases. This has natural monopoly implications if Tesla scales up significantly ahead of the rest of the industry.

 

[neroden]

The video is impressive, but I would like to see Tesla handling a complex urban environment.
I mean... driving in my city in Italy is way more complex than what was shown in that video.

Driving in my rural small town is way more complex even than the crazy San Francisco stuff which I've seen Cruise demo video for. I went through three construction zones, including one where I was being flagged off the road onto a gravel shoulder with very narrow clearance to a ditch, and a one-lane bridge with a blind corner behind it, and debris on the road from a collapsed pile of stones, in one fifteen-minute drive. Oh, and the truck doing K-turn maneuvering to get out of a driveway. There was other weird stuff the day before and yet other weird stuff the day after.

I would love self-driving, because driving is a *pain* here. Is it plausible in the near future? No, it's not.

If they really want to have robo taxi on the streets next year they should already be able to drive in those kind of complex environments.

Yeah, there is precisely zero chance of robotaxis on the streets in general next year (not counting specific geofenced routes).

 

[carterm2]

Intentions...

The investment world thinks EM's intentions for yesterday's presentation was to distract from tomorrows numbers.

Other intentions...

When will cars talk to each other... communicate intentions... reading turn signals is the first step... but what it every car on the road for a mile in any direction was able to network its situation/intentions... platooning would be possible, intersections would become much safer possibly faster. Traffic would improve. Autonomous lanes? Speed limit increases for autonomous cars... everyone would welcome the increase in the capacity of roads.

The Model Y would be a great multi-passenger mover... could you charge more for that? And Why wouldn't you charge more for a ride in an autonomous car? An extra passenger, no small talk, no "fumes". Perhaps if you could set your price you would be better off. Tesla creating a network and charging 25% is a generous offer. Each owner would really become like an Uber/Lyft driver, but only in that they would finance, charge/maintain the car, saving Tesla the expense... basically financing the network for Tesla.

 

[scaesare]

Not sure if it is posted already. Nvidia is using Tesla to sell some chips.


Tesla Raises the Bar for Self-Driving Carmakers | The Official NVIDIA Blog

I think NVIDIA's problem is where they say: " The Xavier processor features a programmable CPU, GPU and deep learning accelerators, delivering 30 TOPs. We built a computer called DRIVE AGX Pegasus based on a two chip solution, pairing Xavier with a powerful GPU to deliver 160 TOPS, and then put two sets of them on the computer, to deliver a total of 320 TOPS"

Those are going to be discrete units with differing capabilities, linked by a bus (or busses), and likely non-unified memory. If that 30 TOPS is 15 TOPS of NN accelerator, 10 TOPS of GPU, and 5 TOPS of CPU, then it's a little disingenuous to lump it all together and say, "Look, 30 TOPS" when the problem at hand needs 25 TOPS worth on NN processing.

Kinda like adding together the horsepower of my truck's ICE engine, starter, and the generator it's towing together and saying "Look a 600 HP solution!".

Interestingly they spent some time talking about workload analysis and design optimization on the presentation...

 

[EVNow]

This discussion reminds me of the vending machine business model. Here's how it works: Buy the vending machine for $2,000 (financing available!) Make a deal to place it in the lobby of a big building. Buy the candy, gum, Fritos, etc for pennies. sell for 1,000% markup! All you need to to so is sell 40 (or some such number) each day x 350 days per year =$Y gross revenue. Subtract your miniscule costs and each machine will make you, say $10,000 per year. All you have to do is restock as needed and collect the money.

Wait -- I have a better idea! Why not buy 5? or 10? What could possibly go wrong?

BTW, we can make the same deal with pinball machines. Oh no, sorry. That ended when municipalities regulated them out of business.

This - but unironically.

This is the exact reason we see vending machines virtually everywhere. The market is saturated.

So, until robotaxi market is saturated - or regulators step in to stop Tesla from expanding, they can keep increasing the size.

BTW, this is one reason I think Google can't actually do this. Regulators will step in very quickly since Google is already a monopoly.

 

[Nocturnal]

You could start by having 2 Model 3s and lock in the FSD price. Add 8 more by the time Tesla Network is up running.

Counterpoint. If all of that is possible, TSLA stock will have a much higher ROI.

 

[pekap]

Quick question for those more knowledgeable. As the fleet grows substantially and higher levels of self driving are reached where human driver needs only intervene when FSD is stumped, once one human driver intervenes and shows how to handle weird corner cases, could this trouble spot be flagged immediately for real-time mimicry of the "lead-car" solution by the Teslas that follow shortly thereafter? Almost like a Waze crowdsource plus a solution, or even just a heads up to the human driver "prepare to take over. Weird situation ahead."

 

[MarcusMaximus]

It isn't like absolutely hard coded for their code. Couldn't be, because that changes. But it isn't just the logic you are referring to, its also the ARM chips, the I/O, the bus width, etc., etc.

I would be very surprised if there was similar efficiency with a dissimilar program/work load. Which is no surprise because it was made to Tesla's specification. Which is how they can out do nvidia for TOPS/watt and even better real performance for their problem per watt.

If you still disagree? Great! Its all theoretical anyway, just a silly tangent brought on by a silly investor observation/question.

I think we’re largely in agreement.

 

[neroden]

Still scratching my head how they freaken made this chip as a side project and suddenly become as competitive as Nvidia while blowing AMD's attempt out of the water. It's really crazy and I don't think these investors really understand what they were looking at.

Yeah, the guy who asked whether they could use the chip for other things was onto something. They should just sell the chips -- make a hell of a lot more money more quickly than the fictional robotaxis are gonna make them.